Measuring Massive Multitask Language Understanding

TL;DR

This paper introduces a comprehensive multitask accuracy test for language models across 57 diverse subjects, revealing significant gaps in current models' knowledge and problem-solving abilities, especially in socially sensitive areas.

Contribution

It presents a new broad evaluation benchmark for measuring language models' multitask understanding across various domains, highlighting current limitations and areas for improvement.

Findings

Largest GPT-3 improves 20 percentage points over random chance

Models perform poorly on socially important subjects like morality and law

Current models still require substantial improvements for expert-level accuracy

Abstract

We propose a new test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more. To attain high accuracy on this test, models must possess extensive world knowledge and problem solving ability. We find that while most recent models have near random-chance accuracy, the very largest GPT-3 model improves over random chance by almost 20 percentage points on average. However, on every one of the 57 tasks, the best models still need substantial improvements before they can reach expert-level accuracy. Models also have lopsided performance and frequently do not know when they are wrong. Worse, they still have near-random accuracy on some socially important subjects such as morality and law. By comprehensively evaluating the breadth and depth of a model's academic and professional understanding, our test can be used to analyze models across many tasks and to identify important shortcomings.